Sentiments analysis of fMRI using automatically generated stimuli labels under naturalistic paradigm

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Linguistics

Background:

• Naturalistic stimuli like movies influence emotions and brain activity.
• Understanding brain activation dynamics aids in identifying neurological conditions such as stress and depression.
• Existing functional magnetic resonance imaging (fMRI) datasets lack emotion/sentiment labels, hindering supervised learning.

Purpose of the Study:

• To develop an automated method for generating emotion/sentiment labels from naturalistic stimuli (movie subtitles).
• To utilize these generated labels for classifying brain fMRI images.
• To improve the utility of open-access fMRI datasets for supervised learning studies.

Main Methods:

• Employed natural language processing sentiment analyzers (VADER, TextBlob, Flair) to extract sentiment labels from movie subtitles.
• Used generated sentiment labels (positive, negative, neutral) as class labels for fMRI data.
• Applied various machine learning classifiers including Support Vector Machine, Random Forest, Decision Tree, and Deep Neural Networks.

Main Results:

• Achieved classification accuracy ranging from 42-84% for imbalanced fMRI datasets.
• Demonstrated improved classification accuracy, ranging from 55-99%, for balanced fMRI datasets.
• Validated the feasibility of using subtitle-derived sentiment labels for fMRI classification.

Conclusions:

• Automated sentiment labeling from subtitles offers a viable alternative to subjective manual labeling for fMRI studies.
• This approach enhances the potential of existing fMRI datasets for supervised learning, advancing the study of emotion and neurological conditions.
• The findings support the integration of natural language processing techniques with neuroimaging for a deeper understanding of brain-emotion interactions.